What E033 Taught Me About Custom AI Assistants

What Building Curio and the Venture Episode Engine Actually Taught Me

In September 2024, I used Gemini Gems and Meta AI Studio to build two versions of Curio, a custom assistant for short facts. I also turned my podcast-outline process into a Gem called the Venture Episode Engine. The experiment made reusable AI feel approachable, but it also showed why a saved configuration is not the same thing as expertise or dependable automation.

The best result was not that a model could write a polished outline. It was that I could move a repeated method out of one fragile chat and into instructions I could inspect, revise, and test.

flowchart LR
    A["Notes and spoken ideas"] --> B["Reusable instructions"]
    B --> C["Draft episode outline"]
    C --> D["Dalton reviews structure"]
    D --> E["Revise instructions"]
    E --> C
    C --> F["Record only after human review"]

Two products exposed two different design instincts

The episode began with a comparison that still matters. Gemini Gems felt like a canvas for persistent instructions. Meta AI Studio felt like a character builder connected to a social network.

I created Curio on both platforms and asked each version about light near a black hole. The exchange was interesting, but it was not a scientific evaluation. It showed how the same concept could be packaged for a private work surface or a public, visual, socially discoverable experience.

Meta's 2024 AI Studio launch record explicitly emphasized custom characters, creator AIs, sharing, and labeled replies. Google's 2024 Gemini Gems announcement described reusable instructions for tasks such as brainstorming, learning, writing, coding, and career help.

That launch contrast gave the episode its shape. It should not be mistaken for a permanent feature comparison. Both products have changed.

The Venture Episode Engine solved a real annoyance

Before Gems, I had used long-running chats to help with podcast outlines. I would explain the structure, give the model notes, and improve the conversation over time. Eventually the behavior would drift. I would struggle to recover the format and start another chat.

The Venture Episode Engine separated the repeatable part from the conversation. Its job was straightforward: take unorganized notes and spoken thoughts, then turn them into the outline structure I used for Venture Step.

That is much closer to a useful assistant than a fictional "world-class expert" persona. I knew the work, the source material, and the expected result. I remained responsible for the episode.

The system still required several rounds of instruction editing. It changed the order of sections. It used much more bold formatting than I wanted. I manually fixed the result. Those small failures were valuable because they were visible and recoverable.

The rewrite button reduced friction, not responsibility

The Gemini interface included an option to rewrite Gem instructions. I gave it rough text and watched it produce a more structured specification. In the live demonstration, that made creation feel quick.

Google's current Gem help page still describes an instruction rewrite feature, along with previewing and optional knowledge files. The feature can be useful as an editor. It cannot know which business rule I forgot, whether a source is authoritative, or whether an action is allowed.

That distinction appeared immediately in the live travel-planner test. I described an upcoming trip to Tokyo and Seoul, a Mount Fuji climb, and a marathon. The assistant created a day-by-day plan with links. It did not know I would be working during the trip because I had not told it.

The output looked organized. The missing constraint still made it unsuitable as a plan.

My broad use of the word agent needs an update

Throughout E033, I called Gems and AI Studio characters "agents." That matched product conversation in 2024, but it blurred several different systems.

The demonstrated Curio configurations stored instructions and responded in chat. The Venture Episode Engine also used persistent instructions, with separate conversations underneath it. None of those examples independently initiated work, carried broad permissions, or executed an open-ended sequence across external systems.

A better description is custom assistant. If a system can use tools, write to other services, continue through a multi-step loop, or act without a new prompt, those capabilities and permissions should be named directly.

The distinction is not vocabulary for its own sake. Authority changes the failure. A poor draft can be revised. An unauthorized message, deleted record, public reply, or exposed file may be harder to recover.

The current product record is different from the episode

Some of the episode's product details aged quickly. I described Gemini Gems as desktop-bound. Google's launch post already said mobile rollout was underway, and current help says a Gem created in the web app can appear in the mobile app and supported Workspace surfaces.

I was also uncertain about AI Studio audience defaults. Current Instagram help describes audience options that can include Everyone, Close Friends, and Only Me, subject to availability and review. The safe lesson is to inspect the actual audience setting, not rely on a remembered default.

Current sharing raises new concerns too. Google's Gem sharing documentation says people with access can view a Gem's instructions and uploaded files. A configuration that feels private can therefore expose more than its creator expects if it is shared.

Those facts belong in a living product record. They should not be written backward into the 2024 recording.

The strongest advice in E033 was to start inward

Near the end, I recommended beginning with a small task that only the user would see, such as organizing notes. That advice has held up better than the feature checklist.

A low-risk internal draft gives the user time to learn what the model misses, what context matters, and how much review the output needs. The next step is not automatically a public assistant or a tool with more authority. It is a better-defined task and a small test set.

NIST's Generative AI Profile frames risk around the specific use case and encourages pre-deployment testing and ongoing evaluation. That is a more durable basis for adoption than the speed of a live demo.

What the episode actually established

E033 established that one creator could turn a recurring outline method into reusable instructions, compare two product designs, and create a rough travel assistant quickly. It also captured instruction drift, formatting problems, a missing travel constraint, ambiguous terminology, and uncertainty about product behavior.

It did not establish autonomous agency, expert performance, measured productivity, safe workplace data use, scientific accuracy, travel reliability, privacy compliance, or public-persona readiness.

That narrower finding is useful. Reusable assistants can save setup time when the task, evidence, permissions, and review are already understood. They do not supply those things on their own.

Continue with [[Gemini Gems and Meta AI Studio Product Record]] for the maintained comparison or [[How to Design a Reusable AI Assistant]] for the method. The travel demonstration leads naturally to [[What My First Days in Tokyo Exposed About My Travel Plan]]. The closing preview connects to [[What Building a Go App With Cursor in Four Hours Actually Proved]] and [[What Venture Step Got Wrong About Reflection 70B]].

Read the existing public episode page, listen on Spotify, or watch the YouTube recording.

This story was developed with AI assistance from the immutable YouTube caption source, the preserved production outline, the legacy article, and linked first-party records. Dalton Anderson remains the author. Transcript, product, privacy, safety, technical, source, and founder review are mandatory before publication. Publication is not authorized.